Abstract
Recommendation system is a new technology to recommend products for customers from huge amounts of products, which infers those objective users' preferences based on their personal information or online behavior. This paper studied the main personalized recommendation technology for current E-commerce. It proposed a hybrid recommendation algorithm based on opinion mining. This system combines web data mining technology, that is, takes advantage of user-generated-content by mining customers' online reviews. It is well known that online reviews can directly reflect customer's real emotion and expectation, so it's appropriate to extract a customer's latent interest and preference from his/her reviews, thus refine recommendation and improve accuracy. Meanwhile, an experiment was conducted and the result demonstrated that our system could generate a reliable and realistic recommendation.
Introduction
In the modern e-commerce landscape, consumers face an overwhelming "dilemma of choice." While traditional recommendation systems—such as content-based filtering and standard collaborative filtering—have mitigated this issue, they often suffer from significant limitations. Content-based systems require rich, accurate metadata that is frequently unavailable, while standard collaborative filtering relies almost exclusively on numerical star ratings.
Numerical ratings are inherently reductive; a "3-star" rating fails to capture why a user was dissatisfied. By ignoring the rich text found in User-Generated Content (UGC), traditional systems miss the latent interests and specific emotional expectations of the customer. This paper proposes a hybrid recommendation algorithm that utilizes opinion mining to extract specific product features and user sentiments from unstructured text, building a more nuanced "User-Feature" model to improve recommendation accuracy.
Proposed Methodology
The proposed system moves beyond simple item-to-item similarity by focusing on the specific attributes users discuss in their reviews. The workflow is divided into several critical stages of data processing and modeling.
Feature Extraction and Bootstrapping
To identify relevant product features (e.g., "battery," "cleanliness," or "price"), the system employs a semi-supervised bootstrapping method. This approach begins with a small set of human-defined "seed words" for various categories. An iterative algorithm then scans the unlabeled review text to discover and incorporate related terms, expanding the feature dictionary dynamically.
The Recommendation Algorithm
The core logic of the system involves isolating specific comments and matching them against the expanded seed lists to determine which features are most important to a given user. The process follows a structured seven-step iterative loop to refine feature labels and calculate word values.
User-Feature (UF) Modeling and Similarity
Instead of a standard User-Item matrix, the system constructs a User-Feature (UF) matrix. In this model, values represent the weight of a user's interest in a specific attribute, based on the frequency of mention. The system then calculates user similarity using Pearson correlation coefficients and Cosine similarity. This allows the system to identify "neighbors" who share specific priorities, such as a preference for high performance over low price.
Experimental Evaluation
Datasets
The researchers validated the algorithm using two distinct datasets:
- TripAdvisor Hotel Reviews: A large-scale set of 235,793 reviews. After filtering for comprehensive reviews covering at least seven key features, a subset of 1,900 hotels and 108,891 reviews was utilized.
- IT168 Mobile Phone Reviews: A qualitative test focusing on four smartphones (Nokia N96, N97, C6, and MOTO XT702) to evaluate multi-objective ranking using PROMETHEE Fuzzy logic.
Results and Discussion
The system's performance was measured using Mean Absolute Error (MAE). The experiments determined that the optimal number of nearest neighbors ($k$) was 5, yielding an MAE of 0.55. Interestingly, as $k$ increased beyond this point, the error rate rose slightly, suggesting that in sparse datasets, a smaller group of highly similar neighbors provides more reliable predictions. In the mobile phone trial, the system’s ranking (MOTO XT702 > Nokia N97 > N96 > C6) aligned perfectly with actual user rankings on the source website, demonstrating real-world applicability.
Conclusion
This research demonstrates that mining latent interests from unstructured text reviews significantly enhances the relevance of personalized recommendations. By treating frequency of mention as a proxy for user interest and employing fuzzy logic for multi-objective decision-making, the system provides a more realistic reflection of consumer behavior. Future work will focus on improving natural language processing to handle sarcasm and reducing the computational overhead required for large-scale text mining.
Summary updated